Re: [PATCH RFC] rcu/kfree: Do not request RCU when not needed

From: Paul E. McKenney
Date: Wed Nov 02 2022 - 16:28:26 EST


On Wed, Nov 02, 2022 at 03:46:59PM -0400, Joel Fernandes wrote:
>
>
> > On Nov 2, 2022, at 2:49 PM, Paul E. McKenney <paulmck@xxxxxxxxxx> wrote:
> >
> > On Wed, Nov 02, 2022 at 07:31:40PM +0100, Uladzislau Rezki wrote:
> >>> On Wed, Nov 02, 2022 at 01:29:17PM -0400, Joel Fernandes wrote:
> >>> On Wed, Nov 2, 2022 at 1:24 PM Uladzislau Rezki <urezki@xxxxxxxxx> wrote:
> >>>>
> >>>> On Wed, Nov 02, 2022 at 09:35:44AM -0700, Paul E. McKenney wrote:
> >>>>> On Wed, Nov 02, 2022 at 12:13:17PM -0400, Joel Fernandes wrote:
> >>>>>> On Wed, Nov 2, 2022 at 8:37 AM Uladzislau Rezki <urezki@xxxxxxxxx> wrote:
> >>>>>>>
> >>>>>>> On Sat, Oct 29, 2022 at 01:28:56PM +0000, Joel Fernandes (Google) wrote:
> >>>>>>>> On ChromeOS, I am (almost) always seeing the optimization trigger.
> >>>>>>>> Tested boot up and trace_printk'ing how often it triggers.
> >>>>>>>>
> >>>>>>>> Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> >>>>>>>> ---
> >>>>>>>> kernel/rcu/tree.c | 18 +++++++++++++++++-
> >>>>>>>> 1 file changed, 17 insertions(+), 1 deletion(-)
> >>>>>>>>
> >>>>>>>> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> >>>>>>>> index 591187b6352e..3e4c50b9fd33 100644
> >>>>>>>> --- a/kernel/rcu/tree.c
> >>>>>>>> +++ b/kernel/rcu/tree.c
> >>>>>>>> @@ -2935,6 +2935,7 @@ struct kfree_rcu_cpu_work {
> >>>>>>>>
> >>>>>>>> /**
> >>>>>>>> * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
> >>>>>>>> + * @rdp: The rdp of the CPU that this kfree_rcu corresponds to.
> >>>>>>>> * @head: List of kfree_rcu() objects not yet waiting for a grace period
> >>>>>>>> * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
> >>>>>>>> * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
> >>>>>>>> @@ -2964,6 +2965,8 @@ struct kfree_rcu_cpu {
> >>>>>>>> struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
> >>>>>>>> raw_spinlock_t lock;
> >>>>>>>> struct delayed_work monitor_work;
> >>>>>>>> + struct rcu_data *rdp;
> >>>>>>>> + unsigned long last_gp_seq;
> >>>>>>>> bool initialized;
> >>>>>>>> int count;
> >>>>>>>>
> >>>>>>>> @@ -3167,6 +3170,7 @@ schedule_delayed_monitor_work(struct kfree_rcu_cpu *krcp)
> >>>>>>>> mod_delayed_work(system_wq, &krcp->monitor_work, delay);
> >>>>>>>> return;
> >>>>>>>> }
> >>>>>>>> + krcp->last_gp_seq = krcp->rdp->gp_seq;
> >>>>>>>> queue_delayed_work(system_wq, &krcp->monitor_work, delay);
> >>>>>>>> }
> >>>>>>>>
> >>>>>>>> @@ -3217,7 +3221,17 @@ static void kfree_rcu_monitor(struct work_struct *work)
> >>>>>>>> // be that the work is in the pending state when
> >>>>>>>> // channels have been detached following by each
> >>>>>>>> // other.
> >>>>>>>> - queue_rcu_work(system_wq, &krwp->rcu_work);
> >>>>>>>> + //
> >>>>>>>> + // NOTE about gp_seq wrap: In case of gp_seq overflow,
> >>>>>>>> + // it is possible for rdp->gp_seq to be less than
> >>>>>>>> + // krcp->last_gp_seq even though a GP might be over. In
> >>>>>>>> + // this rare case, we would just have one extra GP.
> >>>>>>>> + if (krcp->last_gp_seq &&
> >>>>>>>>
> >>>>>>> This check can be eliminated i think. A kfree_rcu_cpu is defined as
> >>>>>>> static so by default the last_gp_set is set to zero.
> >>>>>>
> >>>>>> Ack.
> >>>>>>
> >>>>>>>> @@ -4802,6 +4816,8 @@ static void __init kfree_rcu_batch_init(void)
> >>>>>>>> for_each_possible_cpu(cpu) {
> >>>>>>>> struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
> >>>>>>>>
> >>>>>>>> + krcp->rdp = per_cpu_ptr(&rcu_data, cpu);
> >>>>>>>> + krcp->last_gp_seq = 0;
> >>>>>>>>
> >>>>>>> Yep. This one can be just dropped.
> >>>>>>>
> >>>>>>> But all the rest looks good :) I will give it a try from test point of
> >>>>>>> view. It is interested from the memory footprint point of view.
> >>>>>>
> >>>>>> Ack. Thanks. Even though we should not sample rdp->gp_seq, I think it
> >>>>>> is still worth a test.
> >>>>>
> >>>>> Just for completeness, the main purpose of rdp->gp_seq is to reject
> >>>>> quiescent states that were seen during already-completed grace periods.
> >>>>>
> >>>> So it means that instead of gp_seq reading we should take a snaphshot
> >>>> of the current state:
> >>>>
> >>>> snp = get_state_synchronize_rcu();
> >>>>
> >>>> and later on do a:
> >>>>
> >>>> cond_synchronize_rcu(snp);
> >>>>
> >>>> to wait for a GP.
> >>>
> >>> This can't be called from the timer IRQ handler though (monitor)
> >>>
> >>>> Or if the poll_state_synchronize_rcu(oldstate)) != 0
> >>>> queue_rcu_work().
> >>>
> >>> But something like this should be possible (maybe)
> >>>
> >>>> Sorry for a description using the RCU API functions name :)
> >>>
> >>> I believe you will have to call rcu_poll_gp_seq_start() as well if you
> >>> are using polled API. I am planning to look at this properly more,
> >>> soon. Right now I am going to write up the rcutop doc and share with
> >>> you guys.
> >>>
> >>> (Maybe RCU polling is the right thing to do as we reuse all the infra
> >>> and any corner case it is handling)
> >>>
> >> OK. This is in my todo list also. Since we have discussed it let's move
> >> it forward.
> >>
> >> Below what i have came up with to switch for polling APIs:
> >>
> >>> From 799ce1653d159ef3d35f34a284f738c2c267c75f Mon Sep 17 00:00:00 2001
> >> From: "Uladzislau Rezki (Sony)" <urezki@xxxxxxxxx>
> >> Date: Wed, 2 Nov 2022 19:26:27 +0100
> >> Subject: [PATCH 1/1] rcu: kvfree_rcu: Reduce a memory footptint by using
> >> polling APIs
> >>
> >> Total time taken by all kfree'ers: 6564718459 ns, loops: 10000, batches: 1110, memory footprint: 5057MB
> >> Total time taken by all kfree'ers: 8431051895 ns, loops: 10000, batches: 1109, memory footprint: 2749MB
> >> Total time taken by all kfree'ers: 9477830789 ns, loops: 10000, batches: 1158, memory footprint: 2934MB
> >> Total time taken by all kfree'ers: 9950211144 ns, loops: 10000, batches: 981, memory footprint: 2704MB
> >>
> >> with a patch:
> >>
> >> Total time taken by all kfree'ers: 7712110118 ns, loops: 10000, batches: 1660, memory footprint: 91MB
> >> Total time taken by all kfree'ers: 7002403664 ns, loops: 10000, batches: 1482, memory footprint: 86MB
> >> Total time taken by all kfree'ers: 7842282319 ns, loops: 10000, batches: 1738, memory footprint: 86MB
> >> Total time taken by all kfree'ers: 7230161977 ns, loops: 10000, batches: 1542, memory footprint: 72MB
> >>
> >> Tested with NOCB option, all offloading CPUs:
> >>
> >> kvm.sh --memory 10G --torture rcuscale --allcpus --duration 1 \
> >> --kconfig CONFIG_NR_CPUS=64 \
> >> --kconfig CONFIG_RCU_NOCB_CPU=y \
> >> --kconfig CONFIG_RCU_NOCB_CPU_DEFAULT_ALL=y \
> >> --bootargs "rcuscale.kfree_rcu_test=1 rcuscale.kfree_nthreads=16 \
> >> rcuscale.holdoff=20 rcuscale.kfree_loops=10000 torture.disable_onoff_at_boot" --trust-make
> >>
> >> According to data there is a big gain in memory footprint with a patch.
> >> It is because of call_rcu() and call_rcu_flush() take more effort and
> >> time to queue a callback and then wait for a gp.
> >>
> >> With polling API:
> >> a) we do not need to queue any callback;
> >> b) we might not even need wait for a GP completion.
> >>
> >> Signed-off-by: Uladzislau Rezki (Sony) <urezki@xxxxxxxxx>
> >> ---
> >> kernel/rcu/tree.c | 21 ++++++++++++++++-----
> >> 1 file changed, 16 insertions(+), 5 deletions(-)
> >>
> >> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> >> index 76973d716921..17c3d6f2c55b 100644
> >> --- a/kernel/rcu/tree.c
> >> +++ b/kernel/rcu/tree.c
> >> @@ -2919,18 +2919,20 @@ struct kvfree_rcu_bulk_data {
> >> ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
> >>
> >> /**
> >> + * @rcu_work: A work to reclaim a memory after a grace period
> >> * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
> >> - * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
> >> * @head_free: List of kfree_rcu() objects waiting for a grace period
> >> * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
> >> * @krcp: Pointer to @kfree_rcu_cpu structure
> >> + * @gp_snap: A snapshot of current grace period
> >> */
> >>
> >> struct kfree_rcu_cpu_work {
> >> - struct rcu_work rcu_work;
> >> + struct work_struct rcu_work;
> >> struct rcu_head *head_free;
> >> struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
> >> struct kfree_rcu_cpu *krcp;
> >> + unsigned long gp_snap;
> >> };
> >>
> >> /**
> >> @@ -3066,10 +3068,12 @@ static void kfree_rcu_work(struct work_struct *work)
> >> struct kfree_rcu_cpu_work *krwp;
> >> int i, j;
> >>
> >> - krwp = container_of(to_rcu_work(work),
> >> + krwp = container_of(work,
> >> struct kfree_rcu_cpu_work, rcu_work);
> >> krcp = krwp->krcp;
> >>
> >> + cond_synchronize_rcu(krwp->gp_snap);
> >
> > Might this provoke OOMs in case of callback flooding?
> >
> > An alternative might be something like this:
> >
> > if (!poll_state_synchronize_rcu(krwp->gp_snap)) {
> > queue_rcu_work(system_wq, &krwp->rcu_work);
> > return;
> > }
> >
> > Either way gets you a non-lazy callback in the case where a grace
> > period has not yet elapsed.
> > Or am I missing something that prevents OOMs here?
>
> The memory consumptions appears to be much less in his testing with the onslaught of kfree, which makes OOM probably less likely.
>
> Though, was your reasoning that in case of a grace period not elapsing, we need a non lazy callback queued, so as to make the reclaim happen sooner?
>
> If so, the cond_synchronize_rcu() should already be conditionally queueing non-lazy CB since we don’t make synchronous users wait for seconds. Or did I miss something?

My concern is that the synchronize_rcu() will block a kworker kthread
for some time, and that in callback-flood situations this might slow
things down due to exhausting the supply of kworkers.

In contrast, use of queue_rcu_work() frees up the kworker to handle
other pages that are filling up.

Perhaps your point is that the delay from synchronize_rcu() should make
the following pages take the fastpath through cond_synchronize_rcu()?

Either way, it might well be that context-switch overhead forces us
to batch these things somehow. But let's worry about that when and if
it actually happens.

Thanx, Paul

> Thanks,
>
> - Joel
>
> >
> > Thanx, Paul
> >
> >> +
> >> raw_spin_lock_irqsave(&krcp->lock, flags);
> >> // Channels 1 and 2.
> >> for (i = 0; i < FREE_N_CHANNELS; i++) {
> >> @@ -3194,6 +3198,13 @@ static void kfree_rcu_monitor(struct work_struct *work)
> >> if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) ||
> >> (krcp->bkvhead[1] && !krwp->bkvhead_free[1]) ||
> >> (krcp->head && !krwp->head_free)) {
> >> + /*
> >> + * Take a snapshot for this krwp. Please note no
> >> + * more any objects can be added to this krwp free
> >> + * channels.
> >> + */
> >> + krwp->gp_snap = get_state_synchronize_rcu();
> >> +
> >> // Channel 1 corresponds to the SLAB-pointer bulk path.
> >> // Channel 2 corresponds to vmalloc-pointer bulk path.
> >> for (j = 0; j < FREE_N_CHANNELS; j++) {
> >> @@ -3217,7 +3228,7 @@ static void kfree_rcu_monitor(struct work_struct *work)
> >> // be that the work is in the pending state when
> >> // channels have been detached following by each
> >> // other.
> >> - queue_rcu_work(system_wq, &krwp->rcu_work);
> >> + queue_work(system_wq, &krwp->rcu_work);
> >> }
> >> }
> >>
> >> @@ -4808,7 +4819,7 @@ static void __init kfree_rcu_batch_init(void)
> >> struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
> >>
> >> for (i = 0; i < KFREE_N_BATCHES; i++) {
> >> - INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
> >> + INIT_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
> >> krcp->krw_arr[i].krcp = krcp;
> >> }
> >>
> >> --
> >> 2.30.2
> >>
> >> --
> >> Uladzislau Rezki